Sudoku is graph theory

claude & i embarking on another day of debugging a hopelessly self-destructive GAN

If you run an AI lab, pls make sure your evals team has read this

Are reasoning models better at reasoning or memorization? Our new reasoning benchmark L2 tests R1, o1, o3-mini, GPT 4.5 and Claude 3.7 Sonnet. None of the models performs well but Claude 3.7 Sonnet outperforms o1-preview and R1 by a larger margin!

📣 Introducing LingOly-TOO: A benchmark to separate reasoning from memorisation 📣 Reasoning evals should measure reasoning—and ONLY reasoning! However, current evals are confounded by: 1️⃣ Requiring specific world knowledge 2️⃣ Prior data exposure LingOly-TOO addresses this! 🧵

Don’t be Dense, SLiCE the Cost! 🪓💸 Structured Linear CDEs (SLiCEs) are expressive, efficient, and parallel-in-time sequence models. - ✅ Maximal expressivity without the cost of dense matrices - ⚡ 20× faster per-step training compared to non-linear NCDEs on real-world

I'm attending a Cosmos Institute seminar on the intersection of AI human agency, and collective intelligence, a topic very dear to my heart. Here are some thoughts.

Much love for cogsci-inspired approaches to alignment; well done Sydney Levine and team!

A whole new level of dynamic typing

Such an ambitious project could only be done with the best researchers in the world. Thankfully, the most amazing team has already assembled. Join us! Read the paper at full-stack-alignment.ai/paper

It was terrifically energising to work on this position paper. Floored by the ambition and optimism coming out of the Meaning Alignment Institute team and by the talented cadre they have assembled for this problem. Kudos Ryan Lowe 🥞 @ICML Joe Edelman 🥞 @ICML2025 Oliver Klingefjord 🥞, now the real work begins :)

"What are markets, AI systems, and democratic institutions really for? They are not ends in themselves. We want markets to coordinate human needs and resources. We want democratic institutions to enable collective self-governance. Presumably, we want AI systems to augment human

New from Professor Philipp Koralus (@oxfordhailab) at the The Institute for Ethics in AI: AI that nudges us may quietly undermine autonomy. He proposes a Socratic design for AI – supporting truth-seeking, not steering choices.